Unsupervised Anomaly Detection to Characterize Heterogeneity in Type 2 Diabetes

Area of Science:

• Endocrinology and Metabolism
• Data Science in Healthcare
• Clinical Research

Background:

• Type 2 diabetes presents diverse patient characteristics and clinical pathways.
• Understanding patient heterogeneity is crucial for effective diabetes management.
• Previous research has not fully characterized anomalous patient profiles within large cohorts.

Purpose of the Study:

• To identify anomalous patient characteristics in a large cohort of women with type 2 diabetes.
• To compare the clinical trajectories and treatment patterns of anomalous versus typical patients.
• To inform the development of targeted interventions for high-risk diabetic patients.

Main Methods:

• Utilized dimensionality reduction and anomaly detection techniques.
• Analyzed a large cohort of 21,288 women aged 30-65 with type 2 diabetes.
• Applied preprocessing heuristics to ensure cohort homogeneity in clinical trajectory.

Main Results:

• Anomalous patients were twice as likely to be hospitalized compared to the majority cohort.
• Anomalous patients exhibited a higher incidence of comorbidities (2x more).
• Anomalous patients were prescribed more insulin and fewer newer, expensive medications like SGLT2 inhibitors.

Conclusions:

• Distinct anomalous patient profiles exist within type 2 diabetes cohorts.
• These patients face significantly higher risks for adverse health events.
• Targeted interventions for anomalous patients could mitigate risks and improve outcomes.